Introduction

Project Overview/Problem Statement

Cinderella, Sleeping Beauty, Snow White – classic stories with a tale as old as time. The origins of these and other fairy tales begin long before the Disney versions known today. Written versions of these stories are found in the publications of famous folklorists such as Charles Perrault and the Brothers Grimm [1]. These tales both entertain and intrigue academic interest for textual analysis.

Since fairy tales are tailored to children, they are well suited for text analysis. Vaz et. al. identifies that in comparison to text written for adults, fairy tales have shorter sentences, clearly defined emotions, and a plot and language easily read and understood [2]. Additionally, cultural, generational, and topic differences in fairy tale versions make it possible to research the effect of culture, time, and topic in the use of written sentiment.

Thanks to Project Gutenberg, a volunteer organization that promotes public domain ebooks, 57,000 texts are available to the public free of charge [3]. In this collection are books of Fairy Tales spanning multiple cultures, time periods, and topics. This project is scoped primarily on cultural and topic differences. Fairy Tales with cultural and topic differences were gathered from Household Fairy Tales by the Brothers Grimm, and The Tales of Mother Goose by Charles Perrault. This sample allows for comparison between French and German version of fairy tales, and an assortment of fairy tales with varying topics. Using a corpus built from these texts, this project aims to answer the following three research questions.

  1. Does the sentiment in fairy tales differ based on the topic/ genre of the story?
  2. Does the sentiment in fairy tales vary over the course of the story? If so, does a common pattern of written sentiment emerge over multiple stories?
  3. Does the sentiment in versions of the same fairy tale differ between cultures?

Installing/ Loading Required R Packages

Fifteen packages are used for this project; they are required to follow along with the methodology and reproduce the results. The pacman package is used to facilitate installing and loading these necessary packages. The pacman package should be installed and loaded prior to proceeding.

pacman::p_load(XML,        
               rvest,      
               RCurl,      
               rprojroot,  
               tidytext,   
               stringr,
               pdftools,
               tidyr,
               dplyr,
               yaml,
               ggplot2,
               gutenbergr,
               xts,
               wordcloud,
               reshape2)

Methodology

Word Frequency

In this section word frequency is summarized by term frequency (tf), inverse document frequency (idf), and term inverse document frequency (tf-idf). Term frequency measures the frequency of a word within a document. This is scaled by \(log_{10}\) to downweight the frequency to better measure the relevance of word frequency to the meaning of a document[4]. Thus, term frequency is calculated using the following equation.

\[ tf_{t,d} = \begin{cases} 1+\log_{10}\Big(\mbox{count(t,d)}\Big)\hspace{10pt} \mbox{if count(t,d) > 0}\\\\ 0 \hspace{260px}\mbox{ otherwise } \end{cases} \] Inverse document frequency, is the fraction of the number of documents in a corpus and the number of documents for which a given word appears. This measures how unique a word is to a particular document by reducing the weight for commonly used words and increasing the weight for words that are less frequent in a collection of documents [4]. The equation below shows how idf is calculated.

\[ \mbox{idf}(t,D) = \log\left(\frac{N}{n_{t}}\right) \] Finally, combining tf and idf results in the term inverse document frequency which measures “high frequency words that provide particularly important context to a single document within a group of documents” [9]. This is calculated as the product of tf and idf for a particular term (t) in document (d) for a set of documents (D) and results in a value between 0 (not important) and 1 (very important).

\[ \mbox{tf-idf}(t,d,D) = \mbox{tf}(t,d) \cdot \mbox{idf}(t,D) \]

Sentiment Analysis using tidytext

The three research questions in this project use sentiment analysis. Sentiment Analysis is “the task of extracting the positive or negative orientation that a writer expresses in a text” [4]. In this project sentiment of a text is quantified within the tidyverse ecosystem, where “text [is considered] as a combination of its individual words and the sentiment content of the whole text [is] the sum of the sentiment content of the individual words” [5]. Within the tidy universe, word sentiment can be classified using three different lexicons: AFINN, BING, or NRC. These lexicons are all limited to unigrams and use a dictionary of terms sourced and validated from modern day English. The scoring of sentiment differs between lexicons. The AFINN lexicon ranks sentiment on an integer between -5 (negative) and +5 (positive), the BING lexicon ranks words as -1 (negative), 0 (neutral), or 1 (positive), and the NRC lexicon further subdivides binarily ranked sentiment to categories of “positive, negative, anger, anticipation, disgust, fear, joy, sadness, surprise, and trust” [5]. The following output shows the number of sentiment associated words contained in each lexicon, and the number of sentiment associated words by category within each lexicon.

sentiments %>% 
    group_by(lexicon) %>%
    summarise(noOfWords = n())
# A tibble: 4 x 2
  lexicon  noOfWords
  <chr>        <int>
1 AFINN         2476
2 bing          6788
3 loughran      4149
4 nrc          13901
sentiments %>% 
    filter(lexicon %in% c("nrc", "bing")) %>%
    group_by(lexicon, sentiment) %>%
    summarise(noOfWords = n())
# A tibble: 12 x 3
# Groups:   lexicon [?]
   lexicon sentiment    noOfWords
   <chr>   <chr>            <int>
 1 bing    negative          4782
 2 bing    positive          2006
 3 nrc     anger             1247
 4 nrc     anticipation       839
 5 nrc     disgust           1058
 6 nrc     fear              1476
 7 nrc     joy                689
 8 nrc     negative          3324
 9 nrc     positive          2312
10 nrc     sadness           1191
11 nrc     surprise           534
12 nrc     trust             1231
sentiments %>% 
    filter(lexicon %in% c("AFINN")) %>%
    group_by(lexicon) %>%
    count(cut_width(score, 1))
# A tibble: 11 x 3
# Groups:   lexicon [1]
   lexicon `cut_width(score, 1)`     n
   <chr>   <fct>                 <int>
 1 AFINN   [-5.5,-4.5]              16
 2 AFINN   (-4.5,-3.5]              43
 3 AFINN   (-3.5,-2.5]             264
 4 AFINN   (-2.5,-1.5]             965
 5 AFINN   (-1.5,-0.5]             309
 6 AFINN   (-0.5,0.5]                1
 7 AFINN   (0.5,1.5]               208
 8 AFINN   (1.5,2.5]               448
 9 AFINN   (2.5,3.5]               172
10 AFINN   (3.5,4.5]                45
11 AFINN   (4.5,5.5]                 5

Since the number of sentiment associated words and the quantification of word sentiment differs by lexicon, this project compares sentiment analysis using the three available lexicons to see which works best for comparing sentiment in Fairy Tale’s. Additionally, since the lexicons consist of modern day English terms it is possible that the sentiment of some words in 18th and 19th century folktales is inconsistent or not included in the lexicons used for this project. While this concern is noted, this project proceeds with the three available lexicons and recommends future incorporation of a subject specific sentiment lexicon when developed and made compatable with the tidytext package in R.

Research Questions

Research Question 1: Sentiment Across Topics

The first research question concerns identifying whether sentiment in fairy tales differs across topics. This requires categorizing fairy tales. This project uses the Aarne-Thompson-Uther classification system to distinguish between classes of fairy tales. The topic categories and corresponding indices within each topic are provided in the following table.

\[\begin{array}[] TTopic & Index Range \\ Animal Tales & 1-299 \\ Tales of Magic & 300-749 \\ Religious Tales & 750-849 \\ Realistic Tales & 850-899 \\ Tales of the Stupid Ogre & 1000-1199 \\ Anecdotes and Jokes & 1200-1999 \\ Formula Tales & 2000-2399 \end{array}\]

Each fairy tale was manually indexed based on the classifications provided by [6] and [7]. The code below identifies the titles of fairy tales classified within each category based on the Aarne-Thompson-Uther system. Of the 61 fairy tales in the collection, 3 titles were not assigned to a category. For this project, undesignated fairy tales are excluded when comparing sentiment between topics. The following code identifies the Fairy Tales within each topic.

AnimalTales <- c(" THE WOLF AND THE SEVEN LITTLE GOATS.",
                 " THE WONDERFUL MUSICIAN",
                 " THE STRAW, THE COAL, AND THE BEAN",
                 " THE MOUSE, THE BIRD, AND THE SAUSAGE",
                 " THE BREMEN TOWN MUSICIANS",
                 " HOW MRS FOX MARRIED AGAIN FIRST VERSION",
                 " HOW MRS FOX MARRIED AGAIN SECOND VERSION",
                 " MR KORBES",
                 " OLD SULTAN",
                 " THE DOG AND THE SPARROW"
                 )

TalesOfMagic <- c("CINDERELLA, OR THE LITTLE GLASS SLIPPER.",
                  " THE SLEEPING BEAUTY IN THE WOODS.",
                  " LITTLE THUMB.",
                  " THE MASTER CAT, OR PUSS IN BOOTS.",
                  " RIQUET WITH THE TUFT.",
                  " BLUE BEARD.",
                  " THE FAIRY.",
                  " LITTLE RED RIDING-HOOD.",
                  " SIX SOLDIERS OF FORTUNE",
                  " THE GOOSE GIRL.",
                  " THE RAVEN",
                  " THE FROG PRINCE",
                  " FAITHFUL JOHN",
                  " THE TWELVE BROTHERS",
                  " THE BROTHER AND SISTER",
                  " RAPUNZEL",
                  " THE THREE LITTLE MEN IN THE WOOD",
                  " THE THREE SPINSTERS",
                  " HANSEL AND GRETHEL",
                  " THE WHITE SNAKE",
                  " THE FISHERMAN AND HIS WIFE",
                  " ASCHENPUTTEL",
                  " MOTHER HULDA",
                  " LITTLE RED CAP",
                  " THE TABLE, THE ASS, AND THE STICK.",
                  " TOM THUMB",
                  " THE ELVES",
                  " TOM THUMB'S TRAVELS",
                  " THE ALMOND TREE",
                  " THE SIX SWANS",
                  " THE SLEEPING BEAUTY",
                  " SNOW-WHITE",
                  " THE KNAPSACK, THE HAT, AND THE HORN",
                  " RUMPELSTILTSKIN",
                  " THE GOLDEN BIRD",
                   " THE QUEEN BEE",
                  " THE GOLDEN GOOSE"
                  )

TalesOfTheStupidOgre <- c(" ROLAND")


RealisticTales <- c(" THE ROBBER BRIDEGROOM",
                    " KING THRUSHBEARD"
                    )

AnecdotesAndJokes <- c(" CLEVER GRETHEL",
                       " HANS IN LUCK",
                       " THE GALLANT TAILOR",
                       " CLEVER ELSE",
                       " HOW MRS FOX MARRIED AGAIN FIRST VERSION",
                       " HOW MRS FOX MARRIED AGAIN SECOND VERSION",
                       " FRED AND KATE",
                       " THE LITTLE FARMER"
                       )
                      
FormulaTales <- c(" THE DEATH OF THE HEN")

NoCategoryTales <- c("THE RABBIT'S BRIDE",
                     " THE VAGABONDS",
                     " PRUDENT HANS"
                     )

Research Question 2: Sentiment Throughout A Story

The second research question concerns identifying whether written sentiment in fairy tales vary throughout the course of a story. The motivation for this question is based in Vladimir Propp’s theory of folk tale morphology [8]. Propp, a Russian folklorist, identified 31 narrative units common across folk tales. These 31 narrative chunks were grouped into four spheres: the introduction, the body of the story, the donor sequence, and the hero’s return. While not every story contains all 31 narrative chunks and perhaps the sequence of these chunks is inconsistent between stories, it motivates asking whether written sentiment is different between narrative units and if so, does a common distribution of sentiment throughout fairy tales emerge.

Unfortunately, the corpus of text for this project is not tagged for Propp’s 31 narrative units and it is beyond the scope of this project to attempt to create these tags. Therefore, to show the distribution of sentiment throughout each fairy tale the sentiment is computed for each sentence. Sentence sentiment is quantifed by summing word sentiment within a given sentence and then normalizing the score to account for sentence length. The distribution of sentiment in a story can then be visualized as a time series of normalized sentence sentiment scores.

Research Question 3: Sentiment Across Cultures

The final research question concerns identifying whether written sentiment varies between cultural versions of fairy tales. As a case study, versions of Cinderella, Sleeping Beauty, and Little Red Riding Hood are compared between French and German authors, Perrault and the Brothers Grimm.

Data

The gutenbergr package by David Robinson provides access to the Project Gutenberg collection from within R. In this project, the works from the Brothers Grimm and Charles Perrault are accessed from the function gutenberg_download() that downloads each eBook by referencing the Project Gutenberg ID.

grimm <- gutenberg_download(19068)
Perrault <- gutenberg_download(17208)

After uploading the raw data, irrelevant text that is not part of the fairy tales must be removed. This includes removing introductory text at the start of the book, concluding text at the end of the book, and illustration placeholders. Additionally, it is necessary to format the text so that we can easily identify the titles of each fairy tale. The following code prepares the raw data to create the corpus of text for this project.

grimm <- grimm %>%
  mutate(gutenberg_id = "Brothers Grimm")%>%
  filter(row_number()>238)%>% #Remove text at beginning of book
  filter(row_number()<10806)%>% #Remove text at end of book
  filter(!str_detect(text, regex("^    \\[Illustration")))%>% #Remove Illustration Placeholders
  filter(!str_detect(text, regex("^\\[Illust"))) #Remove Illustration Placeholders 

replace_grimm6569 ="HOW MRS FOX MARRIED AGAIN FIRST VERSION" #Fix the name of a title
replace_grimm6637 ="HOW MRS FOX MARRIED AGAIN SECOND VERSION" #Fix the name of a title 

#Make lowercase non title lines that are capitalized
#Make uppercase title lines that are not capitalized
grimm$text <- c(grimm$text[1:382],toupper(grimm$text[383]),
                grimm$text[384:705],tolower(grimm$text[706:709]),grimm$text[710:1451],
                toupper(grimm$text[1452]), grimm$text[1453:1576], tolower(grimm$text[1577:1583]),
                grimm$text[1584:2707],tolower(grimm$text[2708:2709]),grimm$text[2710:2839], 
                toupper(grimm$text[2840]),grimm$text[2841:3744],toupper(grimm$text[3745]),
                grimm$text[3746:3803],toupper(grimm$text[3804]),
                grimm$text[3805:4868],toupper(grimm$text[4869]),
                grimm$text[4870:4955],tolower(grimm$text[4956:4961]), grimm$text[4962:5844],
                toupper(grimm$text[5845]), grimm$text[5846:6568],replace_grimm6569,
                grimm$text[6570:6636], replace_grimm6637, grimm$text[6638:6834],
                tolower(grimm$text[6835:6837]),grimm$text[6838:7242], 
                tolower(grimm$text[7243:7244]), grimm$text[7245:7768],
                tolower(grimm$text[7769:7773]),grimm$text[7774:8277], 
                tolower(grimm$text[8278:8281]),grimm$text[8282:8677],
                toupper(grimm$text[8678]),grimm$text[8679:9241],
                tolower(grimm$text[9242:9244]),grimm$text[9245:9535],
                toupper(grimm$text[9536]),grimm$text[9537:9695],
                toupper(grimm$text[9696]),grimm$text[9697:10589])
  
Perrault <- Perrault %>%
  mutate(gutenberg_id = "Perrault")%>%
  filter(row_number()>139)%>% #Remove text at beginning of book
  filter(row_number()<1867)%>% #Remove text at end of book
  filter(!str_detect(text, regex("^    \\[Illustration")))%>%  #Remove Illustration Placeholders
  filter(!str_detect(text, regex("^\\[Illust"))) #Remove Illustration Placeholders

After cleaning the data, for each eBook, the raw text is formatted as a data frame in which each observation is a line of text from the eBook. In total, there are 61 fairy tales that comprise the corpus for this project. However, in its original format, the text of each book is not separated by story. Nevertheless, preceding each story is a fully capitalized title as shown below for the first fairy tale in Household Fairy Tales by the Brothers Grimm. Fairy tale titles are similarly capitalized in the text for Charles Perrault.

grimm[1:6,2]
# A tibble: 6 x 1
  text                                                                   
  <chr>                                                                  
1 THE RABBIT'S BRIDE                                                     
2 ""                                                                     
3 ""                                                                     
4 THERE was once a woman who lived with her daughter in a beautiful      
5 cabbage-garden; and there came a rabbit and ate up all the cabbages. At
6 last said the woman to her daughter,

Utilizing regex, the start of each story is identified from capitalization and thereby each line of text is assigned to its corresponding story. Additionally, the cleaned text is transformed into a tidy data structure by making each word an observation. The sentence and paragraph in which each word appears is stored as additional variables in the dataset. This preserves the structure of the original text in varying granularity for the second research question which quantifies the written sentiment across the duration of each story. Additionally, each story is indexed according to the Aarne-Thompson-Uther classification system used for the first research question. Finally, the corpus for this project is completed by combining the fairy tales of the Brothers Grimm and Charles Perault into a single data frame. The following code executes these steps.

FairyTaleCorpus <- bind_rows(grimm,Perrault) #Combine fairy tales into a single corpus
FairyTaleCorpus <- FairyTaleCorpus[-c(47608:47617),]

FairyTaleCorpus <- FairyTaleCorpus %>% 
  unnest_tokens(paragraph,text,token = "paragraphs", to_lower = FALSE) %>% #Unnest text into paragraphs
  mutate(story_title = ifelse(str_detect(paragraph, regex("^[12\\,\\.\\[\\]A-Z \\'\\-]+$")), paragraph, NA),
         story = na.locf(story_title)) %>% # Indicate the starting location of each story
  filter(is.na(story_title)) %>% # Remove fairy tale titles from the text
  select(-story_title) %>% # Remove the variable story_title 
  mutate(story_index = row_number())%>%
  mutate(p_index = row_number())%>%
  unnest_tokens(sentence,paragraph,token = "sentences", drop = FALSE, collapse=FALSE) %>% #Unnest text into sentences
  mutate(s_index = row_number())%>%
  unnest_tokens(word,sentence,token = "words", drop = FALSE, collapse=FALSE) %>% #Unnest text into words
  mutate(class = ifelse(story %in% AnimalTales, "Animal Tales", # Index each story type
                        ifelse(story %in% TalesOfMagic, "Tales Of Magic",
                               ifelse(story %in% RealisticTales, "Realistic Tales",
                                      ifelse(story %in% TalesOfTheStupidOgre, "Tales Of The Stupid Ogre",
                                        ifelse(story %in% AnecdotesAndJokes, "Anecdotes And Jokes",
                                             ifelse(story %in% FormulaTales, "Formula Tales","No Category")))))))

### This code is used to fix the paragraph and sentence index so it restarts for each story.

p_ref = 0 # Initialize paragraph reference at 0
s_ref = 0 # Initialize sentence reference at 0
story_ref = 1

for (i in 2:nrow(FairyTaleCorpus)){ # For each observation...
  if(FairyTaleCorpus$gutenberg_id[i]!=FairyTaleCorpus$gutenberg_id[i-1]){
    pref = 1
    s_ref = 0
  }
  if(FairyTaleCorpus$story[i]!=FairyTaleCorpus$story[i-1]){ # If you start a new story... 
    story_ref = story_ref + 1
    p_ref = FairyTaleCorpus$p_index[i]-1 # Mark the last paragraph index of the previous story
    s_ref = FairyTaleCorpus$s_index[i]-1 # Mark the last sentence index of the previous story
  }
  
  FairyTaleCorpus$story_index[i]=story_ref # Scale paragraph index for story
  FairyTaleCorpus$p_index[i]=FairyTaleCorpus$p_index[i]-p_ref # Scale paragraph index for story
  FairyTaleCorpus$s_index[i]=FairyTaleCorpus$s_index[i]-s_ref # Scale sentence index for story
}

FairyTaleCorpus <- FairyTaleCorpus %>%
  group_by(sentence) %>%
  mutate(SentenceTotal = sum(s_index) %/% s_index)%>%
  ungroup()

After completing the aforementioned steps, the corpus of fairy tales, in tidy format, is ready for sentiment analysis.

Analysis

Word Frequency

Word frequency across authors, topics, and stories are explored prior to conducting sentiment analysis. This intermediate step helps to become acquainted to the data and ensure it was properly cleaned.

The following code calculates tf, idf, and tf-idf for the words within the corpus and visualizes the results using the ggplot2 package. The visualization of word frequency is done across authors, topics, and fairy tales in accordance to the three levels of research questions that this project addresses.

Word Frequency By Author

The following word clouds show the words which occur most frequently by author according to the tf and tf-idf scores respectively. When comparing frequent words by authors it appears as though the tidytext package did an adequate job removing stop words in spite of differences in language used in the era’s of these fairy tales and today. In the tf word cloud we see some instances of frequently used words that are common between authors such king, wife, and time. Words which are frequent but not common between authors are shown in the tf-idf word cloud and include author specific words such as hans and ogre. It appears that frequent words less common between authors are lots of character names or types. Interestingly, the word faithful is used frequently by the Brothers Grimm but not Perrault which might appear later when comparing sentiment across authors/ cultures. In general, the term frequency results are consistent with words that one might expect to appear frequently in fairy tales.

author_words %>%
   dplyr::arrange(dplyr::desc(tf)) %>%
        dplyr::mutate(word = base::factor(word, levels = base::rev(base::unique(word))),
               author = base::factor(gutenberg_id)) %>% 
        dplyr::group_by(author) %>%
        dplyr::top_n(15, wt = tf) %>%
  acast(word ~ gutenberg_id, value.var = "tf", fill = 0) %>%
  comparison.cloud(colors = c("#F8766D", "#00BFC4"),title.bg.colors = c("#F8766D", "#00BFC4"),
                   max.words = 100)

author_words %>%
   dplyr::arrange(dplyr::desc(tf_idf)) %>%
        dplyr::mutate(word = base::factor(word, levels = base::rev(base::unique(word))),
               author = base::factor(gutenberg_id)) %>% 
        dplyr::group_by(author) %>%
        dplyr::top_n(15, wt = tf_idf) %>%
  acast(word ~ gutenberg_id, value.var = "tf_idf", fill = 0) %>%
  comparison.cloud(colors = c("#F8766D", "#00BFC4"),title.bg.colors = c("#F8766D", "#00BFC4"),
                   max.words = 100)

Word Frequency By Topic

The following word clouds show the words which occur most frequently by topic according to the tf and tf-idf scores respectively. These plots continue to confirm that the tidytext package did an adequate job removing stop words. Additionally, words that intuitevely align to topics appear under each category. For instance, in the tf wordcloud under the Animal Tales category we see words such as wolf, dog, and cat appear and in Tales of Magic we see words appear such as princess and king. Nevertheless, other categories overlap, also showing high term frequency of a subset of these words. In the tf-idf wordcloud we see some of the same words appear as the tf wordcloud but we also see other words appear which are likely more unique to a certain topic. It is interesting to note that the term ogre is not frequent for tf or tf-idf in the Tales of the Stupid Ogre category even though it appeared frequent when calculating tf-idf word frequency across authors.

mypalette<-brewer.pal(9,"Dark2")

topic_words %>%
   dplyr::arrange(dplyr::desc(tf)) %>%
        dplyr::mutate(word = base::factor(word, levels = base::rev(base::unique(word))),
               topic = base::factor(class)) %>% 
        dplyr::group_by(topic) %>%
        dplyr::top_n(15, wt = tf) %>%
  acast(word ~ topic, value.var = "tf", fill = 0) %>%
  comparison.cloud(colors = mypalette,title.bg.colors = mypalette,
                   max.words = 100)

topic_words %>%
   dplyr::arrange(dplyr::desc(tf_idf)) %>%
        dplyr::mutate(word = base::factor(word, levels = base::rev(base::unique(word))),
               topic = base::factor(class)) %>% 
        dplyr::group_by(topic) %>%
        dplyr::top_n(15, wt = tf_idf) %>%
  acast(word ~ topic, value.var = "tf_idf", fill = 0) %>%
  comparison.cloud(colors = mypalette,title.bg.colors = mypalette,
                   max.words = 100)

Word Frequency By Story

Although tf, idf and tf-idf was calculated for all 61 stories, to visualize word frequency by story we will only plot the versions of Cinderella, Sleeping Beauty and Little Red Riding-Hood since these stories are analyzed in greater depth being the case study for both sentiment across stories and sentiment between cultures/ authors. Thus, the following word clouds show words across six stories (the two versions of the three fairy tales) with the highest tf and tf-idf scores respectively.

Word frequency by fairy tales across this sample identifies some potential issues in the data. For starters, in the tf-idf wordcloud the word grandmother appears for the Brothers Grimm version of Little Red Riding-Hood while grandmamma and grandmother’s appears for Perrault’s version of the tale. These variations all refer to grandmother, however they are not identified as equivalent. Another insight is the common words unique to each version of the fairy tale Cinderella. In general, it appears that word frequency between stories for these three tales did not perform as well as word frequency between authors and topics. Variations of frequent words appear as unique across stories when in reality they are not unique. This project notes this concern but does not assume the tedious task of manually fixing these word variations. Standardizing variations of the same word is noted as a future research extension which would likely improve the accuracy of results.

mypalette<-brewer.pal(6,"Dark2")

story_words %>%
  filter(story %in% c(" LITTLE RED RIDING-HOOD."," LITTLE RED CAP",
                      "CINDERELLA, OR THE LITTLE GLASS SLIPPER.", " ASCHENPUTTEL",
                      " THE SLEEPING BEAUTY IN THE WOODS.", " THE SLEEPING BEAUTY")) %>%
   dplyr::arrange(dplyr::desc(tf)) %>%
        dplyr::mutate(word = base::factor(word, levels = base::rev(base::unique(word))),
               story = base::factor(story)) %>% 
        dplyr::group_by(story) %>%
        dplyr::top_n(15, wt = tf) %>%
  acast(word ~ story, value.var = "tf", fill = 0) %>%
  comparison.cloud(colors = mypalette,title.bg.colors = mypalette,
                   max.words = 100)

story_words %>%
  filter(story %in% c(" LITTLE RED RIDING-HOOD."," LITTLE RED CAP",
                      "CINDERELLA, OR THE LITTLE GLASS SLIPPER.", " ASCHENPUTTEL",
                      " THE SLEEPING BEAUTY IN THE WOODS.", " THE SLEEPING BEAUTY")) %>%
   dplyr::arrange(dplyr::desc(tf_idf)) %>%
        dplyr::mutate(word = base::factor(word, levels = base::rev(base::unique(word))),
               story = base::factor(story)) %>% 
        dplyr::group_by(story) %>%
        dplyr::top_n(15, wt = tf_idf) %>%
  acast(word ~ story, value.var = "tf_idf", fill = 0) %>%
  comparison.cloud(colors = mypalette,title.bg.colors = mypalette,
                   max.words = 100)

Research Question 1: Sentiment by Topic

While preparing the data, fairy tales were classified according to their Aarne-Thompson-Uther Index. Prior to using this information to find sentiment across topics it is worth noting the number of fairy tales and words within each category. The distribution of fairy tales into each of the 7 categories is summarized in the table below.

\[\begin{array}[] TTopic & Number of Fairy Tales \\ Animal Tales & 10 \\ Tales of Magic & 37 \\ Religious Tales & 0 \\ Realistic Tales & 2 \\ Tales of the Stupid Ogre & 1 \\ Anecdotes and Jokes & 9 \\ Formula Tales & 1 \\ Not Assigned & 3 \\ \end{array}\]

The following output shows the number of words within each topic.

table(FairyTaleCorpus$class)

     Anecdotes And Jokes             Animal Tales            Formula Tales 
                   12450                     7779                      629 
             No Category          Realistic Tales           Tales Of Magic 
                    2321                     3017                    78069 
Tales Of The Stupid Ogre 
                    1451

There is a disparity between the number of fairy tales and words within each topic. This sample bias could cause errors when comparing sentiment across Aarne-Thompson-Uther topics. To mitigate some of the bias, sentiment scores by topic will be standardized to account for the total number of words within each classification.

Sentiment BING Lexicon

The following code assigns word sentiment to text using the BING lexicon. The sentiment by topic is quantified as the proportion of positive and negative words which appear to the total number of words within each topic. The proportion of sentiment by topic is plotted using the ggplot2 package.

bing_word_counts_byTopic <- FairyTaleCorpus %>%
  filter(!class %in% "No Category") %>%
  inner_join(get_sentiments("bing"))%>%
  inner_join(get_sentiments("bing")) %>%
  group_by(class) %>%
  count(word, sentiment, sort = TRUE) %>%
  ungroup()%>%
  left_join(topic_total_words)%>%
  mutate(sentiment_proportion = n/total)

bing_word_counts_byTopic %>%   
  group_by(class, sentiment)%>%
  summarise(sentiment_proportion = sum(sentiment_proportion)) %>%
  ggplot(aes(sentiment, sentiment_proportion, fill = class)) +
  geom_col(show.legend = FALSE) +
  facet_wrap(~class, scales = "free_y") +
  labs(y = "Contribution to sentiment",
       x = NULL) +
  coord_flip()

From the graph it appears that using BING sentiment lexicons, the Formula Tales category are proportionally the most negative and least positive. Animal Tales also has a higher distribution of negative sentiment than positive sentiment. Realistic Tales, Tales Of Magic, Anecdotes And Jokes and Tales of the Stupid Ogre all appear to have a higher proportion of positive sentiment than negative sentiment. Tales of Magic, Tales of the Stupid Ogre, and Anecdote and Joke categories have similarly distributed positive and negative sentiment proportions.

We can validate these results by viewing how each word contributed to the positive and negative sentiment score across topics. The following code generates a plot to view the top 10 words which had the most significant positive and negative contribution to sentiment. From the plot it appears that the BING sentiment lexicon does fairly well for quantifying sentiment across topics. The words that contribute to positive and negative sentiment are sound and there is not any apparant issues.

bing_word_counts_byTopic %>%
  group_by(sentiment,class) %>%
  top_n(10) %>%
  ungroup() %>%  
  mutate(n = ifelse(sentiment == "negative", -n, n)) %>%
  mutate(word = reorder(word, n)) %>%
   ggplot(aes(word, n/total, fill = sentiment)) +
  geom_col(show.legend = TRUE) +
  facet_wrap(~class, scales = "free_y") +
  labs(y = "Contribution to sentiment",
       x = NULL) +
  coord_flip()

AFINN Lexicon

The following code assigns word sentiment to text using the AFINN lexicon. The sentiment by topic is quantified as the proportion of words in each factor of ranked sentiment, from -5 (negative) to 5 (positive), with respect to the total number of words within each topic. The proportion of sentiment by topic is plotted using the ggplot2 package.

afinn_word_counts_byTopic <- FairyTaleCorpus %>%
  filter(!class %in% "No Category") %>%
  inner_join(get_sentiments("afinn"))%>%
  group_by(class) %>%
  count(word, score, sort = TRUE) %>%
  ungroup()%>%
  left_join(topic_total_words)%>%
  mutate(score_proportion = n/total)

afinn_word_counts_byTopic %>%   
  group_by(class, score)%>%
  summarise(score_proportion = sum(score_proportion)) %>%
  ggplot(aes(score, score_proportion, fill = class)) +
  geom_col(show.legend = FALSE) +
  facet_wrap(~class, scales = "free_y") +
  labs(y = "Contribution to sentiment",
       x = NULL) +
  coord_flip()

From the graph it appears that using AFINN sentiment lexicons, the Formula Tales category have a larger proportion of strongly ranked negative sentiment compared to the other categories. Using the word contribution to sentiment scores we can see why this occurs.

afinn_word_counts_byTopic %>%
  group_by(score,class) %>%
  top_n(10) %>%
  ungroup() %>%  
  mutate(n = ifelse(score <0, -n, n)) %>%
  mutate(word = reorder(word, n)) %>%
  ggplot(aes(word, n/total, fill = score)) +
  geom_col(show.legend = TRUE) +
  facet_wrap(~class, scales = "free_y") +
  labs(y = "Contribution to sentiment",
       x = NULL) +
  coord_flip()

It appears that the words ass and cock, which refers to a donkey and rooster respectively, are not recognized by the AFINN lexicon as being animals and so these word are being ranked highly negative and biasing the results for the Formula Tales topic. Removing instances of these words will provide a more accurate representation of sentiment scores using the AFINN lexicon.

afinn_word_counts_byTopic <- FairyTaleCorpus %>%
  filter(!class %in% "No Category") %>%
  filter(!word %in% c("ass","cock"))%>%
  inner_join(get_sentiments("afinn"))%>%
  group_by(class) %>%
  count(word, score, sort = TRUE) %>%
  ungroup()%>%
  left_join(topic_total_words)%>%
  mutate(score_proportion = n/total)

afinn_word_counts_byTopic %>%   
  group_by(class, score)%>%
  summarise(score_proportion = sum(score_proportion)) %>%
  ggplot(aes(score, score_proportion, fill = class)) +
  geom_col(show.legend = FALSE) +
  facet_wrap(~class, scales = "free_y") +
  labs(y = "Contribution to sentiment",
       x = NULL) +
  coord_flip()

Removing these words removed the category of -5 negative words within Formula Tales and provides a more accurate representation of sentiment across topics using the AFINN lexicon. While it is nice that the AFINN lexicon attempts to quantify positivity or negativity of words beyond binary classification, it is likely that the scoring of sentiment is highly subjective and could bias results for words that appear in a different context than the lexicon was trained. This should be noted when using the AFINN lexicon in future research questions.

NRC Lexicon

The following code assigns word sentiment to text using the NRC lexicon. The word sentiment is classified as either anticipation, joy, positive, surprise, trust, sadness, negative, disgust, fear, and/or anger. To compare the written sentiment across topics, the proportion of words by topic within each sentiment category is calculated.

nrc_word_counts_byTopic <- FairyTaleCorpus %>%
  filter(!class %in% "No Category") %>%
  filter(!word %in% c("cock","ass")) %>%
  inner_join(get_sentiments("nrc"))%>%
  group_by(class) %>%
  count(word, sentiment, sort = TRUE) %>%
  ungroup()%>%
  mutate(sentiment = reorder(sentiment, n)) %>%
  left_join(topic_total_words)

nrc_word_counts_byTopic %>%
  ggplot(aes(sentiment, n/total, fill=class))+
  geom_col(show.legend = FALSE)+
  facet_wrap(~class, ncol = 3, scales = "free")+
  coord_flip()

From the plot above we see different proportions for expressed sentiment across topics but similar trends. For all topics the NRC lexicon finds a higher proportion of positive than negative words. Additionally, it appears that the Animal Tales category has the highest proportion of negative sentiment. The following code generates a plot to see which words contribute to the proportions of each sentiment by category.

nrc_word_counts_byTopic %>%
  group_by(sentiment,class) %>%
  top_n(10) %>% 
  mutate(n = ifelse(sentiment %in% c("sadness","negative","disgust",
                                     "fear","anger"), -n, n)) %>%
  mutate(word = reorder(word, n)) %>%
  ggplot(aes(word, n/total, fill = sentiment)) +
  geom_col(show.legend = TRUE) +
  facet_wrap(~class, scales = "free_y") +
  labs(y = "Contribution to sentiment",
       x = NULL) +
  coord_flip()

From the graph it appears that words can be categorized into multiple sentiment categories. Since the sentiment of a word is likely context dependent it is possible that further subdivision of word sentiment beyond positive and negative can also introduce more bias. Despite the subjectivity and possible bias, the ability to determine emotions in fairy tales by categorizing sentiment beyond a binary classification seems useful. Thus, if using this lexicon to categorize sentiment extra caution should be taken to limit the introduction of additional subjectivity and bias.

RESEARCH QUESTION 2: Sentiment Throughout Story

In the first research question sentiment was aggregated across topics. Alternatively, we can aggregate sentiment over sentences within a story. By aggregating sentiment over sentences we should be able to identify critical moments of emotion throughout a story. Ideally, this analysis will reveal a common trend or pattern that exists across fairy tales. While it is possible to run this analysis for all 61 stories, a subset of six (three fairy tales of two authors), is used as case study. The fairy tales selected are each authors versions of Cinderella, Snow White, and Little Red Riding Hood. Viewing this subset of fairy tales will both provide insight towards sentiment change throughout a story and help towards identifying differences in versions of the fairy tales between authors/ cultures.

BING Lexicon

The following code aggregates sentiment across sentences using the BING, AFINN, and NRC lexicons and standardizes it based on the number of words in each sentence. For the NRC lexicons only the positive and negative sentiment categories are used.

#Calculate Bing Sentiment
story_sentiment_bing <- FairyTaleCorpus %>%
  filter(story %in% c(" LITTLE RED RIDING-HOOD."," LITTLE RED CAP",
                      "CINDERELLA, OR THE LITTLE GLASS SLIPPER.", " ASCHENPUTTEL",
                      " THE SLEEPING BEAUTY IN THE WOODS.", " THE SLEEPING BEAUTY")) %>%
  inner_join(get_sentiments("bing"))%>%
  count(story, index = s_index, sentiment, SentenceTotal, sentence) %>%
  spread(sentiment, n, fill = 0) %>%
  mutate(sentiment = (positive-negative)/SentenceTotal) %>%
  mutate(method = "bing")

story_sentiment_bing <- story_sentiment_bing[,-(5:6)]

#Calculate AFINN Sentiment
story_sentiment_afinn <- FairyTaleCorpus %>%
  filter(story %in% c(" LITTLE RED RIDING-HOOD."," LITTLE RED CAP",
                      "CINDERELLA, OR THE LITTLE GLASS SLIPPER.", " ASCHENPUTTEL",
                      " THE SLEEPING BEAUTY IN THE WOODS.", " THE SLEEPING BEAUTY")) %>%
  inner_join(get_sentiments("afinn"))%>%
  count(story, index = s_index, score, SentenceTotal, sentence)

story_sentiment_afinn <- story_sentiment_afinn %>%
  group_by(story,index, SentenceTotal, sentence)%>%
  summarize(sentiment = sum(score))%>%
  mutate(sentiment = sentiment/SentenceTotal)%>%
  ungroup()%>%
  mutate(method = "afinn")

# Calculate NRC Sentiment
story_sentiment_nrc <- FairyTaleCorpus %>%
  filter(story %in% c(" LITTLE RED RIDING-HOOD."," LITTLE RED CAP",
                      "CINDERELLA, OR THE LITTLE GLASS SLIPPER.", " ASCHENPUTTEL",
                      " THE SLEEPING BEAUTY IN THE WOODS.", " THE SLEEPING BEAUTY")) %>%
  inner_join(get_sentiments("nrc"))%>%
  filter(sentiment %in% c("positive","negative"))%>%
  count(story, index = s_index, sentiment, SentenceTotal, sentence) %>%
  spread(sentiment, n, fill = 0) %>%
  mutate(sentiment = (positive-negative)/SentenceTotal) %>%
  mutate(method = "nrc")

story_sentiment_nrc <- story_sentiment_nrc[,-(5:6)]
x <- rbind(story_sentiment_bing,story_sentiment_afinn, story_sentiment_nrc)


ggplot(data = x, mapping = aes(x = index, y = sentiment, fill = method)) +
    geom_bar(alpha = 0.8, stat = "identity", show.legend = FALSE) +
    facet_wrap(facets = ~ story+method, ncol = 3, scales = "free_x")

Generally the sentiment trends throughout stories seems consistent across lexicons. Neverthless, it appears that the AFINN lexicon may be inflating the sentiment scores of some sentences compared to the bing and nrc lexicons. To view which sentences the sentiment scores are being inflated we can view the top scoring positive and negative sentences with the AFINN lexicon.

The following code sorts the sentence sentiment and outputs the six most negative sentences according to the AFINN lexicon. It appears that the AFINN lexicon is biasing the results towards short sentences such as “cried the grandmother” and “oh no!”.

story_sentiment_afinn %>%
  group_by(method)%>%
  arrange(sentiment)%>%
  head()
# A tibble: 6 x 6
# Groups:   method [1]
  story         index SentenceTotal sentence              sentiment method
  <chr>         <dbl>         <dbl> <chr>                     <dbl> <chr> 
1 " LITTLE RED~    26             3 cried the grandmothe~    -0.667 afinn 
2 " ASCHENPUTT~   115             2 "\"oh no!"               -0.5   afinn 
3 " THE SLEEPI~    97             4 cried the head cook.     -0.5   afinn 
4 " ASCHENPUTT~     7            11 and then began very ~    -0.455 afinn 
5 " THE SLEEPI~    10            20 "\"in the fifteenth ~    -0.4   afinn 
6 CINDERELLA, ~    95             8 these by no means an~    -0.375 afinn

The same issue occurs with positive sentences. The top 3 scoring sentences from the afinn method is just the exclamation “ha!” Based on these results I do not recommend the AFINN lexicon for calculating sentiment across sentences in fairy tales.

story_sentiment_afinn %>%
  arrange(desc(sentiment))%>%
  head()
# A tibble: 6 x 6
  story                index SentenceTotal sentence       sentiment method
  <chr>                <dbl>         <dbl> <chr>              <dbl> <chr> 
1 " THE SLEEPING BEAU~    27             1 "\"ha!\""           2    afinn 
2 CINDERELLA, OR THE ~    60             1 "\"ha!"             2    afinn 
3 CINDERELLA, OR THE ~    62             1 ha!                 2    afinn 
4 " ASCHENPUTTEL"         19             2 "\"fine cloth~      1    afinn 
5 CINDERELLA, OR THE ~    61             4 how beautiful~      0.75 afinn 
6 CINDERELLA, OR THE ~    63             4 "how beautifu~      0.75 afinn

The BING and NRC lexicons are far superior for scoring negative sentiment across sentences. The top negative results more accurately reflect critical moments of emotion within the text.

story_sentiment_bing %>%
  arrange(sentiment)%>%
  head()
# A tibble: 6 x 6
  story         index SentenceTotal sentence              sentiment method
  <chr>         <dbl>         <dbl> <chr>                     <dbl> <chr> 
1 CINDERELLA, ~     4            14 the wedding was scar~    -0.214 bing  
2 " ASCHENPUTT~   130            20 and so they were con~    -0.2   bing  
3 " ASCHENPUTT~     7            11 and then began very ~    -0.182 bing  
4 " THE SLEEPI~   109            17 now the poor chief c~    -0.176 bing  
5 " ASCHENPUTT~    17            12 and as she always lo~    -0.167 bing  
6 " LITTLE RED~    66             6 but the grandmother ~    -0.167 bing  
story_sentiment_nrc %>%
  arrange(sentiment)%>%
  head()
# A tibble: 6 x 6
  story         index SentenceTotal sentence              sentiment method
  <chr>         <dbl>         <dbl> <chr>                     <dbl> <chr> 
1 " LITTLE RED~    48            18 and, saying these wo~    -0.222 nrc   
2 " LITTLE RED~    44             5 "\"the better to dev~    -0.2   nrc   
3 " ASCHENPUTT~     8            12 "\"is the stupid cre~    -0.167 nrc   
4 " THE SLEEPI~    16            13 this terrible gift m~    -0.154 nrc   
5 CINDERELLA, ~     4            14 the wedding was scar~    -0.143 nrc   
6 " LITTLE RED~    29            23 little red riding-ho~    -0.130 nrc

Reviewing the highest scoring positive words it appears that BING outperformed both the AFINN and NRC lexicons. The NRC lexicon does better than the AFINN lexicon for classifying positive sentiment but it still appears that short sentences without significant meaning are biased.

story_sentiment_bing %>%
  arrange(desc(sentiment))%>%
  head()
# A tibble: 6 x 6
  story            index SentenceTotal sentence           sentiment method
  <chr>            <dbl>         <dbl> <chr>                  <dbl> <chr> 
1 " ASCHENPUTTEL"     19             2 "\"fine clothes!\~     0.5   bing  
2 " LITTLE RED CA~    10             6 "\"thank you kind~     0.333 bing  
3 CINDERELLA, OR ~    61             4 how beautiful she~     0.25  bing  
4 CINDERELLA, OR ~    63             4 "how beautiful sh~     0.25  bing  
5 " LITTLE RED RI~    39             9 "\"that is the be~     0.222 bing  
6 " ASCHENPUTTEL"    121             5 "\"this is the ri~     0.2   bing  
story_sentiment_nrc %>%
  arrange(desc(sentiment))%>%
  head()
# A tibble: 6 x 6
  story          index SentenceTotal sentence             sentiment method
  <chr>          <dbl>         <dbl> <chr>                    <dbl> <chr> 
1 " ASCHENPUTTE~     9             9 "said they; \"those~     0.333 nrc   
2 " LITTLE RED ~    26             3 cried the grandmoth~     0.333 nrc   
3 " LITTLE RED ~    20             3 called the grandmot~     0.333 nrc   
4 " THE SLEEPIN~    76             7 the prince helped t~     0.286 nrc   
5 " THE SLEEPIN~    94            11 "\"i intend to eat ~     0.273 nrc   
6 " ASCHENPUTTE~    37             4 "you want to dance!~     0.25  nrc

From visual inspection it seems like quantifying sentiment across sentences does fairly well to show how sentiment changes throughout a story but there are not any distinct trends that appear across multiple stories.

Interestingly, different versions of the same fairy tale end with different sentiment. Identifying the last sentence of each fairy tale can validate whether this insight has significance.

story_sentiment_bing %>%
  group_by(story) %>%
  filter(index == max(index)) %>%
  arrange(story)
# A tibble: 6 x 6
# Groups:   story [6]
  story         index SentenceTotal sentence              sentiment method
  <chr>         <dbl>         <dbl> <chr>                     <dbl> <chr> 
1 " ASCHENPUTT~   130            20 and so they were con~   -0.2    bing  
2 " LITTLE RED~    70            12 then little red-cap ~   -0.0833 bing  
3 " LITTLE RED~    48            18 and, saying these wo~   -0.111  bing  
4 " THE SLEEPI~    42            23 then the wedding of ~    0.0435 bing  
5 " THE SLEEPI~   124            25 the king was of cour~    0.04   bing  
6 CINDERELLA, ~   120            32 cinderella, who was ~    0.0938 bing

From the last sentences of these six stories we can easily identify the ending sentiment of the fairy tale and how it differs across versions of the same fairy tale. This leads to the third research question concerning how sentiment varies across cultural versions of the same story.

RESEARCH QUESTION 3: Sentiment Across Author/ Cultural Versions of Fairy Tales

From the second research question we identified that in Cinderella, Little Red Riding Hood and Sleeping Beauty the sentiment of the last sentence differed between versions. Interestingly, the version of the fairy tale with the most positive sentiment of the final sentence aligns with the version of the story that was adapted by Disney (i.e. Walt Disney modeled Sleeping Beauty from the Brothers Grimm version and Cinderella from Charles Perrault’s version [10] and both of these versions had a final sentence with more positive sentiment then their counterparts.) This follows intuition since Disney fairy tales are known for having happy endings. From this insight, we conjecture that the concluding sentences of a fairy tale can be a good representation for the overall sentiment of a story.

To further explore sentiment of fairy tales across authors/ cultures we quantify the overall sentiment of each story for the Brothers Grimm and Perrault. Given the comparison of the three lexicons from the first two research questions we limit this analysis to the BING lexicon which was found to be more accurate than the other two lexicons.

From this analysis it appears as though a similar pattern in the proportion of positive and negative sentiment occurs between versions of Little Red Riding Hood. However, there is a difference in the sentiment proportions across versions of Cinderella and Sleeping Beauty. As expected, Charles Perrault’s version of Cinderella is more positive than the Brothers’ Grimm version. In Sleeping Beauty the proportion of sentiment is contrary to what was expected. The Brothers Grimm version (which was used as a template for Disney’s version and had more positive sentiment in the final sentence) is overall more negative and less positive than the version by Perrault.

Quantifying the sentiment by author we see that overall, the Brothers Grimm fairy tales tend to be more negative than Charles Perrault’s fairy tales which reconfirms previous findings.

Findings and Conclusions

Overall sentiment analysis for fairy tales proved fruitful. This project seperated text by Aarne-Thomspon-Uther categories, sentences within stories, and by author to explore how sentiment in fairy tales vary across topics, cultures, and throughout a story. Three lexicons, AFINN, BING, and NRC were compared for assigning word sentiment. The results from the three research questions indicate that the BING sentiment lexicon produced the most accurate results. Sentiment analysis across topics identified that the proportion of sentiment between topics only varies slightly and that many topics have similar trends between positive and negative sentiment. Plotting the distribution of sentiment across sentences we found that the BING lexicon worked best and that we were able to identify important sentences and moments in the story with regard to emotion. Even though a common trend across stories was not seen we were able to see how sentiment varied throughout a story and found that the sentiment in a final sentence of a fairy tale differs between cultural versions. Finally, the sentiment analysis across cultures showed that the Brother’s Grimm tend to be more negative than Charles Perrault and that this was apparant throughout the cultural versions of Cinderella and Sleeping Beauty but not Little Red Riding Hood.

Future Work

There is ample opportunity to expand upon the research from this project. Based on the findings from this project we recommend the following extensions:

  • Incorporate topic specific sentiment lexicons that are more accurate for 18th and 19th century fairy tales.
  • Consider an alternative methodology for quantifying sentiment that is not dependent on the use of lexicons.
  • Increase the sample of text in the corpus
  • Compare sentence sentiment to document summmarization to see how well the absolute value of standardized sentiment counts can highlight key sentences.

References